Projmake-mode: Flymake Replacement

There is a great new Emacs plugin from Eric Merritt that like FlyMake builds your code and highlights within Emacs any errors or warnings, but unlike FlyMake builds across the whole project. You can clone the mode from here projmake-mode

After cloning the repo to your desired location add this bit to your dot emacs file, replacing <PATH> with the path to where you cloned the repo.

This Emacs code also uses add-hook to set projmake-mode to start for erlang-mode is loaded. Projmake by default knows how to handle rebar and Make based builds so there is no setup after this, assuming your project is built this way.

Here is my Makefile for building Erlang projects with rebar, replace PROJECT with the name of your project:

Now you can load Emacs and a file from your project and if it is an Erlang file due to the add-hook function in our dot emacs file it will automatically load projmake-mode. You can add hooks for other modes or simply run M-x projmake-mode.

For more documentation and how to extend to other types of projects check out the documentation.


Sinan Releases and Being Right

Fred, of Learn You Some Erlang for Great Good, today posted on his blog about the problems around how rebar handles releases, Rebar Releases and Being Wrong. The problems he mentions and a few others are why, despite giving it a legitimate shot, I have found rebar unusable for my workflow to be efficient and stable while adhering to OTP standards at the same time.

I suggest first reading his post, if you already use rebar, and then continuing on with the rest of this.

I’ll start with an example on the generation of a project containing two applications and a dependency from one of those applications of cowboy. Next, I’ll create a release (and in the process a deployable target system) to show the difference in how sinan handles this process.

TL;DR Sinan does OTP the right way, rebar does not.

First, you can download the latest version sinan from this link, it is simply an executable escript, so ‘chmod +x sinan‘ and put it in your PATH and you are good to go.

Sinan provides a ‘gen’ command to create your project. I include the output of the steps I took to build this project. Sinan assumes this is a multiple application project, but if you give “y” instead it will create a directory structure similar to rebars default structure with a src/ directory instead of a lib/ directory.

$ sinan gen
Please specify your name 
your name> Tristan Sloughter
Please specify your email address 
your email>
Please specify the copyright holder 
copyright holder ("Tristan Sloughter")> 
Please specify name of your project
project name> rel_example
Please specify version of your project
project version> 0.0.1
Please specify the ERTS version ("5.9")> 
Is this a single application project ("n")> 
Please specify the names of the OTP apps that will be developed under this project. One application to a line. Finish with a blank line.
app> app_1
app ("")> app_2
app ("")> 
Would you like a build config? ("y")> y
Project was created, you should be good to go!

We now have a project named rel_example and can see the generated contents.

$ cd rel_example/
$ ls
config  doc  lib  sinan.config

Before going further I add the line {include_erts, true}. to sinan.config so that a generated tarball of the release contains erts and can be booted on a machine without Erlang installed.

$ cat sinan.config
{project_name, rel_example}.
{project_vsn, "0.0.1"}.

{build_dir,  "_build"}.

{ignore_dirs, ["_", "."]}.

{ignore_apps, []}.

{include_erts, true}.

A tree structure view of the generated project is below:

├── config
│   └── sys.config
├── doc
├── lib
│   ├── app_1
│   │   ├── doc
│   │   ├── ebin
│   │   │   └── overview.edoc
│   │   ├── include
│   │   └── src
│   │       ├── app_1_app.erl
│   │       ├──
│   │       └── app_1_sup.erl
│   └── app_2
│       ├── doc
│       ├── ebin
│       │   └── overview.edoc
│       ├── include
│       └── src
│           ├── app_2_app.erl
│           ├──
│           └── app_2_sup.erl
└── sinan.config

You’ll see we have a lib directory with two applications containing their source files under a src directory. Now in order to boot the release we’ll create, we need to remove a couple tings from each supervisor. Instead of creating something for them to supervise just remove the variable AChild and replace [AChild] with [].

Next, so we have a third party dependency in the example, add cowboy to the applications in nano lib/app_1/src/

{applications, [kernel, stdlib, cowboy]},

Sinan provides a depends command to show the depenedencies of the project and where they are located:

$ sinan depends -v
starting: depends
Using the following lib directories to show resolved dependencies and where it found them:


compile time dependencies:

runtime dependencies:

    kernel                    2.15       : /home/tristan/.kerl/installs/r15b/lib/kernel-2.15
    stdlib                    1.18       : /home/tristan/.kerl/installs/r15b/lib/stdlib-1.18
    cowboy                    0.5.0      : /home/tristan/.kerl/installs/r15b/lib/cowboy-0.5.0

project applications:

    app_1                     0.1.0      : /home/tristan/Devel/rel_example/_build/rel_example/lib/app_1-0.1.0
    app_2                     0.1.0      : /home/tristan/Devel/rel_example/_build/rel_example/lib/app_2-0.1.0

Now lets build a release and target system.

$ sinan dist

After running the dist command we have a _build directory that we find the following structure. I removed the files/dirs under each app to shorten the listing.

├── rel_example
│   ├── bin
│   │   ├── rel_example
│   │   └── rel_example-0.0.1
│   ├── erts-5.9
│   │   ├── 
│   ├── lib
│   │   ├── app_1-0.1.0
│   │   │   ├── 
│   │   ├── app_2-0.1.0
│   │   │   ├── 
│   │   ├── cowboy-0.5.0
│   │   │   ├── 
│   │   ├── kernel-2.15
│   │   │   ├── 
│   │   └── stdlib-1.18
│   │       ├── 
│   └── releases
│       └── 0.0.1
│           ├── rel_example.boot
│           ├── rel_example.rel
│           ├── rel_example.script
│           └── sys.config
└── tar
    └── rel_example-0.0.1.tar.gz

Sinan has created a lib directory containing all necessary applications for our release as well as the needed files for booting the release. Additionally the dist command creates a tar.gz for easy deployment. But if we simply want to run our release where we are we can:

$ _build/rel_example/bin/rel_example
Erlang R15B (erts-5.9)

[64-bit] [smp:4:4] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.9  (abort with ^G)

This is only the tip of the iceberg of what sinan is capable of. I can’t go into all of it here but I’ll mention that you are able to define multiple releases for a project to generate and which of your project apps to include in each. Additionally you are able to provide a custom rel file if you require tweaks.

The important part to take away from this post is the structure of what you are working with when using sinan and how it is based on OTP standards, both for the source you work on and the results of the build process under _build/.

Property based testing for unit testers with PropEr – Part 1

This tutorial is brought to you by ErlangCamp 2011 (click here) – Boston, August 12th and 13th – It’s gonna be totally sweet!

Main contributors: Torben Hoffmann, Raghav Karol, Eric Merritt

The purpose of the short document is to help people who are familiar
with unit testing understand how property based testing (PBT) differs,
but also where the thinking is the same.

This document focusses on the PBT tool
PropEr for Erlang since that is
what I am familiar with, but the general principles applies to all PBT
tools regardless of which language they are written in.

The approach taken here is that we hear from people who are used to
working with unit testing regarding how they think when designing
their tests and how a concrete test might look.

These descriptions are then “converted” into the way it works with
PBT, with a clear focus on what stays the same and what is different.

Testing philosophies

A quote from Martin Logan (@martinjlogan):

For me unit testing is about contracts. I think about the same things
I think about when I write statements like {ok, Resp} =
Mod:Func(Args). Unit testing and writing specs are very close for me.
Hypothetically speaking lets say a function should return return {ok,
string()} | {error, term()} for all given input parameters then my
unit tests should be able to show that for a representative set of
input parameters that those contracts are honored. The art comes in
thinking about what that set is.

The trap in writing all your own tests can often be that we think
about the set in terms of what we coded for and not what may indeed be
asked of our function. As the code is tried in further exploratory
testing and in production new input parameter sets for which the given
function does not meet the stated contract are discovered and added to
the test case once a fix has been put into place.

This is a very good description of what the ground rules for unit
testing are:

  • Checking that contracts are obeyed.
  • Creating a representative set of input parameters.

The former is very much part of PBT – each property you write will
check a contract, so that thinking is the same.

xUnit vs PBT

Unit testing has become popular for software testing with the advent
of xUnit tools like jUnit for Java. xUnit like tools typically
provide a testing framework with the following functionality

  • test fixture setup
  • test case execution
  • test fixture teardown
  • test suite management
  • test status reporting and management

While xUnit tools provide a lot of functionality to execute and manage
test cases and suites, reporting results there is no focus on test
case execution step, while this is the main focus area of
property-based testing (PBT).

Consider the following function specification

sort(list::integer()) ---> list::integer() | error

A verbal specification of this function is,

For all input lists of integers, the sort function returns a sorted
list of integers.

For any other kind of argument the function returns the atom error.

The specification above may be a requirement of how the function
should behave or even how the function does behave. This distinction
is important; the former is the requirement for the function, the
latter is the actual API. Both should be the same and that is what our
testing should confirm. Test cases for this function might look like

assertEqual(sort([5,4,3,2,1]), [1,2,3,4,5])
assertEqual(sort([1,2,3,4,5]), [1,2,3,4,5])
assertEqual(sort([]         ), []         )
assertEqual(sort([-1,0, 1]  ), [-1, 0, 1] )

How many tests cases should we write to be convinced that the actual
behaviour of the function is the same as its specification? Clearly,
it is impossible to write tests cases for all possible input values,
here all lists of integers, the art of testing is finding individual
input values that are representative of a large part of the input
space. We hope that the test cases are exhaustive to cover the
specification. xUnit tools offer no support for this and this is where
PBT and PBT Tools like PropEr and QuickCheck come in.

PBT introduces testing with a large set of random input values and
verifying that the specification holds for each input value
selected. Functions used to generate input values, generators, are
specified using rules and can be simply composed together to construct
complicated values. So, a property based test for the function above
may look like:

FOREACH({I, J, InputList},  {nat(), nat(), integer_list()},
    SUCHTHAT(I < J andalso J < length(InputList),
    SortedList = sort(InputList)
    length(SortedList) == length(InputList)
    lists:get(SortedList, I) =< lists:get(SortedList, J))

The property above works as follows

  • Generate a random list of integers InputList and two natural numbers
    I, J, such that I < J < size of InputList
  • Check that size of sorted and input lists is the same.
  • Check that element with smaller index I is less than or equal to
    element with larger index J in SortedList.

Notice in the property above, we specify property. Verification of
the property based on random input values will be done by the property
based tool, therefore we can generated a large number of tests cases
with random input values and have a higher level of confidence that
the function when using unit tests alone.

But it does not stop at generation of input parameters. If you have
more complex tests where you have to generate a series of events and
keep track of some state then your PBT tool will generate random
sequences of events which corresponds to legal sequences of events and
test that your system behaves correctly for all sequences.

So when you have written a property with associated generators you
have in fact created something that can create numerous test cases –
you just have to tell your PBT tool how many test cases you want to
check the property on.

Shrinking the bar

At this point you might still have the feeling that introducing the
notion of some sort of generators to your unit testing tool of choice
would bring you on par with PBT tools, but wait there is more to

When a PBT tool creates a test case that fails there is real chance
that it has created a long test case or some big input parameters –
trying to debug that is very much like receiving a humongous log from
a system in the field and try to figure out what cause the system to

Enter shrinking…

When a test case fails the PBT tool will try to shrink the failing
test case down to the essentials by stripping out input elements or
events that does not cause the failure. In most cases this results in
a very short counterexample that clearly states which events and
inputs are required to break a property.

As we go through some concrete examples later the effects of shrinking
will be shown.

Shrinking makes it a lot easier to debug problems and is as key to the
strength of PBT as the generators.

Converting a unit test

We will now take a look at one possible way of translating a unit
test into a PBT setting.

The example comes from Eric Merritt and is about the add/2 function in
the ec_dictionary instance ec_gb_trees.

The add function has the following spec:

-spec add(ec_dictionary:key(), ec_dictionary:value(), Object::dictionary()) ->

and it is supposed to do the obvious: add the key and value pair to
the dictionary and return a new dictionary.

Eric states his basic expectations as follows:

  1. I can put arbitrary terms into the dictionary as keys
  2. I can put arbitrary terms into the dictionary as values
  3. When I put a value in the dictionary by a key, I can retrieve that same value
  4. When I put a different value in the dictionary by key it does not change other key value pairs.
  5. When I update a value the new value in available by the new key
  6. When a value does not exist a not found exception is created

The first two expectations regarding being able to use arbritrary
terms as keys and values is a job for generators.

The latter four are prime candidates for properties and we will create
one for each of them.


key() -> any().

value() -> any().

For PropEr this approach has the drawback that creation and shrinking
becomes rather time consuming, so it might be better to narrow to
something like this:

key() -> union([integer(),atom()]).

value() -> union([integer(),atom(),binary(),boolean(),string()]).

What is best depends on the situation and intended usage.

Now, being able to generate keys and values is not enough. You also
have to tell PropEr how to create a dictionary and in this case we
will use a symbolic generator (detail to be explained later).

sym_dict() ->

sym_dict(0) ->
sym_dict(N) ->
                  {1, {'$call',ec_dictionary,remove,[key(),sym_dict(N-1)]}},
                  {2, {'$call',ec_dictionary,add,[value(),value(),sym_dict(N-1)]}}

sym_dict/0 uses the ?SIZED macro to control the size of the
generated dictionary. PropEr will start out with small numbers and
gradually raise it.

sym_dict/1 is building a dictionary by randomly adding key/value
pairs and removing keys. Eventually the base case is reached which
will create an empty dictionary.

The ?LAZY macro is used to defer the calculation of the
sym_dict(N-1) until they are needed and frequency/1 is used
to ensure that twice as many adds compared to removes are done. This
should give rather more interesting dictionaries in the long run, if
not one can alter the frequencies accondingly.

But does it really work?

That is a good question and one that should always be asked when
looking at genetors. Fortunately there is a way to see what a
generator produces provided that the generator functions are exported.

Hint: in most cases it will not hurt to throw in a
-compile(export_all). in the module used to specify the
properties. And here we actually have a sub-hint: specify the
properties in a separate file to avoid peeking inside the
implementation! Base the test on the published API as this is what the
users of the code will be restricted to.

When the test module has been loaded you can test the generators by
starting up an Erlang shell (this example uses the erlware_commons
code so get yourself a clone to play with):

$ erl -pz ebin -pz test
1> proper_gen:pick(ec_dictionary_proper:key()).
2> proper_gen:pick(ec_dictionary_proper:key()).
3> proper_gen:pick(ec_dictionary_proper:key()).
4> proper_gen:pick(ec_dictionary_proper:key()).
5> proper_gen:pick(ec_dictionary_proper:key()).
{ok,'36\207_là ´?\nc'}
6> proper_gen:pick(ec_dictionary_proper:value()).
7> proper_gen:pick(ec_dictionary_proper:value()).
8> proper_gen:pick(ec_dictionary_proper:value()).
9> proper_gen:pick(ec_dictionary_proper:value()).
10> proper_gen:pick(ec_dictionary_proper:value()).
11> proper_gen:pick(ec_dictionary_proper:value()).
12> proper_gen:pick(ec_dictionary_proper:value()).
13> proper_gen:pick(ec_dictionary_proper:value()).
14> proper_gen:pick(ec_dictionary_proper:value()).
15> proper_gen:pick(ec_dictionary_proper:sym_dict()).
16> proper_gen:pick(ec_dictionary_proper:sym_dict()).

That does not look too bad, so we will continue with that for now.

Properties of add/2

The first expectation Eric had about how the dictionary works was that
if a key had been stored it could be retrieved.

One way of expressing this could be with this property:

prop_get_after_add_returns_correct_value() ->
    ?FORALL({Dict,K,V}, {sym_dict(),key(),value()},
             try ec_dictionary:get(K,ec_dictionary:add(K,V,Dict)) of
                    V ->
                    _ ->
                   _:_ ->

This property reads that for all dictionaries get/2 using a key
from a key/value pair just inserted using the add/3 function
will return that value. If that is not the case the property will
evaluate to false.

Running the property is done using proper:quickcheck/1:

OK: Passed 100 test(s).

This was as expected, but at this point we will take a little detour
and introduce a mistake in the ec_gb_trees implementation and see
how that works.

Signatures – Reusable, Toolable, Testable Types

This tutorial is brought to you by ErlangCamp 2011 – Boston, August 12th and 13th – It’s gonna be totally sweet!

It often occurs in coding that we need a library, a set of functionality. Often there are several algorithms that could provide
this functionality. However, the code that uses it, either doesn’t care about the individual algorithm or wishes to delegate choosing that algorithm to some higher level. Lets take the concrete example of dictionaries. A dictionary provides the ability to access a value via a key (other things as well but primarily this). There are may ways to implement a dictionary. Just a few are:

Each of these approaches has their own performance characteristics,  memory footprints etc. For example, a table of size n with open addressing has no collisions and holds up to n elements, with a single comparison for successful lookup, and a table of size n with chaining and k keys has the minimum max(0, k-n) collisions and O(1 + k/n) comparisons for lookup.  For skip lists the performance characteristics are about as good as that of randomly built binary search trees – namely (O log n). So the choice of which to select depends very much on memory available, insert/read characteristics, etc. So delegating the choice to a single point in your code is a very good idea. Unfortunately, in Erlang that’s not so easy to do at the moment.

Other languages, have built-in support for this functionality. Java has InterfacesSML has Signatures. Erlang, though, doesn’t currently support this model, at least not directly. There are a few ways you can approximate it. One way is to pass the Module name to the calling functions along with the data that it is going to be called on.

add(ModuleToUse, Key, Value, DictData) ->
    ModuleToUse:add(Key, Value, DictData).

This works, and you can vary how you want to pass the data. For example, you could easily use a tuple to contain the data. That is, you could pass in {ModuleToUse, DictData} and that would make it a bit cleaner.

add(Key, Value, {ModuleToUse, DictData}) ->
    ModuleToUse:add(Key, Value, DictData).

Either way, there are a few problems with this approach. One of the biggest is that you lose code locality, by looking at this bit of code you don’t know what ModuleToUse is at all. You would need to follow the call chain up to figure out what it is. Also it may not be obvious what is actually happening. The fact that ModuleToUse is a variable name obscures the code making it harder to understand. The other big problem is that the tools provided with Erlang can’t help find mistakes that you might have made. Tools like Xref and Dialyzer have just as hard a time figuring out the what ModuleToUse is pointing to as you do. So they can’t give you warnings about potential problems. In fact someone could inadvertently pass an unexpected function name as ModuleToUse and you would never get any warnings, just an exception at run time.

Fortunately, Erlang is a pretty flexible language so we can use a similar approach with a few adjustments to give us the best of both worlds. Both the flexibility of ignoring a specific implementation and keeping all the nice locality we get by using an explicit module name.

So what we actually want to do is something mole like this:

add(Key, Value, DictData) ->
    dictionary:add(Key, Value, DictData).

Doing this we retain the locality. We can easily look up the dictionary Module. We immediately have a good idea what a dictionary actually is and we know what functions we are calling. Also, all the tools know what a dictionary is as well and how to check that your code is calling it correctly. For all of these reasons, this is a much better approach to the problem. This is what Signatures are all about.


How do we actually do this in Erlang when Erlang is missing what Java, SML and friends has built-in?

The first thing we need to do is to define a Behaviour for our functionality. To continue our example we will define a Behaviour for dictionaries. That Behaviour looks like this:



behaviour_info(callbacks) ->
    [{new, 0},
     {has_key, 2},
     {get, 2},
     {add, 3},
     {remove, 2},
     {has_value, 2},
     {size, 1},
     {to_list, 1},
     {from_list, 1},
     {keys, 1}];
behaviour_info(_) ->

So we have our Behaviour now. Unfortunately, this doesn’t give us much yet. It will make sure that any dictionaries we write will have all the functions they need, but it wont help us actually use those dictionaries in an abstract way in our code. To do that we need to add a bit of functionality. We do that by actually implementing our own behaviour, starting with new/1.

%% @doc create a new dictionary object from the specified module. The
%% module should implement the dictionary behaviour.
%% @param ModuleName The module name.
-spec new(module()) -> dictionary(_K, _V).
new(ModuleName) when is_atom(ModuleName) ->
    #dict_t{callback = ModuleName, data = ModuleName:new()}.

This code creates a new dictionary for us. Or to be more specific it actually creates a new dictionary Signature record, that will be used subsequently in other calls. This might look a bit familiar from our previous less optimal approach. We have both the module name and the data. here in the record. We call the module name named in ModuleName to create the initial data. We then construct the record and return that record to the caller and we have a new dictionary. What about the other functions; the ones that don’t create a dictionary but make use of it. Let’s take a look at the implementations of two kinds of functions, one that updates the dictionary and another that just retrieves data.

The first we will look at is the one that updates the dictionary by adding a value.

%% @doc add a new value to the existing dictionary. Return a new
%% dictionary containing the value.
%% @param Dict the dictionary object to add too
%% @param Key the key to add
%% @param Value the value to add
-spec add(key(K), value(V), dictionary(K, V)) -> dictionary(K, V).
add(Key, Value, #dict_t{callback = Mod, data = Data} = Dict) ->
    Dict#dict_t{data = Mod:add(Key, Value, Data)}.

There are two key things here.

  1. The dictionary is deconstructed so we can get access to the data
    and the callback module.
  2. We modify the dictionary record with the new data and return that
    modified record.

This is the same approach that you will use for any Signature that updates data. As a side note, notice that we are calling the concrete implementation to do the work itself.

Now let’s do a data retrieval function. In this case, the get function of the dictionary Signature.

%% @doc given a key return that key from the dictionary. If the key is
%% not found throw a 'not_found' exception.
%% @param Dict The dictionary object to return the value from
%% @param Key The key requested
%% @throws not_found when the key does not exist
-spec get(key(K), dictionary(K, V)) -> value(V).
get(Key, #dict_t{callback = Mod, data = Data}) ->
    Mod:get(Key, Data).

In this case, you can see a very similar approach to deconstructing the dict record. We still need to pull out the callback module and the data itself and call the concrete implementation of the algorithm. In this case, we return the data returned from the call, not the record itself.

That is really all you need to define a Signature. There is a complete implementation in erlware_commons/ec_dictionary.

Using Signatures

It’s a good idea to work through an example so we have a bit better idea of how to use these Signatures. If you are like me, you probably have some questions about what kind of performance burden this places on the code. At the very least we have an additional function call along with the record deconstruction. This must add some overhead. So lets write a little timing test, so we can get a good idea of how much this is all costing us.

In general, there are two kinds of concrete implementations for Signatures. The first is a native implementations, the second is a

Native Signature Implementations

A Native Signature Implementation is just that, a module that implements the Behaviour defined by a Signature directly. For most user defined Signatures this is going to be the norm. In our current example, the erlware_commons/ec_rbdict module is the best example of a Native Signature Implementation. It implements the ec_dictionary module directly.

Signature Wrappers

A Signature Wrapper is a module that wraps another module. Its purpose is to help a pre-existing module implement the Behaviour defined by a Signature. A good example of this in our current example is the erlware_commons/ec_dict module. It implements the ec_dictionary Behaviour, but all the functionality is provided by the stdlib/dict module itself. Lets take a look at one example to see how this is done.

We will take a look at one of the functions we have already seen. The get function an ec_dictionary get doesn’t have quite the same semantics as any of the functions in the dict module. So a bit of translation needs to be done. We do that in the ec_dict module get function.

-spec get(ec_dictionary:key(K), Object::dictionary(K, V)) ->
get(Key, Data) ->
    case dict:find(Key, Data) of
    {ok, Value} ->
     error ->

So the ec_dict module’s purpose for existence is to help the preexisting dict module implement the Behaviour defined by the Signature.

Why do we bring this up here? Because we are going to be looking at timings, and Signature Wrappers add an extra level of indirection to the mix and that adds a bit of additional overhead.

Creating the Timing Module

We are going to create timings for both Native Signature Implementations and Signature Wrappers.

Lets get started by looking at some helper functions. We want dictionaries to have a bit of data in them. So to that end we will create a couple of functions that create dictionaries for each type we want to test. The first we want to time is the Signature Wrapper, so dict vs ec_dict called as a Signature.

create_dict() ->
lists:foldl(fun(El, Dict) ->
        dict:store(El, El, Dict)
    end, dict:new(),

The only thing we do here is create a sequence of numbers 1 to 100, and then add each of those to the dict as an entry. We aren’t too worried about replicating real data in the dictionary. We care about timing the function call overhead of Signatures, not the performance of the dictionaries themselves.

We need to create a similar function for our Signature based dictionary ec_dict.

create_dictionary(Type) ->
lists:foldl(fun(El, Dict) ->
        ec_dictionary:add(El, El, Dict)

Here we actually create everything using the Signature. So we don’t need one function for each type. We can have one function that can create anything that implements the Signature. That is the magic of Signatures. Otherwise, this does the exact same thing as the dict create_dict/1.

We are going to use two function calls in our timing. One that updates data and one that returns data, just to get good coverage. For our dictionaries we are going to use the size function as well as the add function.

time_direct_vs_signature_dict() ->
    io:format("Timing dict~n"),
    Dict = create_dict(),
    test_avg(fun() ->
             dict:size(dict:store(some_key, some_value, Dict))
    io:format("Timing ec_dict implementation of ec_dictionary~n"),

The test_avg function runs the provided function the number of times specified in the second argument and collects timing information. We are going to run these one million times to get a good average (its fast so it doesn’t take long). You can see that in the anonymous function that we directly call dict:size/1 and dict:store/3 to perform the test. However, because we are in the wonderful world of Signatures we don’t have to hard code the calls for the Signature implementations. Lets take a look at the time_dict_type function.

time_dict_type(Type) ->
    io:format("Testing ~p~n", [Type]),
    Dict = create_dictionary(Type),
    test_avg(fun() ->
         ec_dictionary:size(ec_dictionary:add(some_key, some_value, Dict))

As you can see we take the type as an argument (we need it for dict creation) and call our create function. Then we run the same timings that we did for ec dict. In this case though, the type of dictionary is never specified, we only ever call ec_dictionary, so this test will work for anything that implements that Signature.

dict vs ec_dict Results

So we have our tests, what was the result. Well on my laptop this is what it looked like.

Erlang R14B01 (erts-5.8.2) [source] [64-bit] [smp:4:4] [rq:4] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.8.2  (abort with ^G)

1> ec_timing:time_direct_vs_signature_dict().
Timing dict
Range: 2 - 5621 mics
Median: 3 mics
Average: 3 mics
Timing ec_dict implementation of ec_dictionary
Testing ec_dict
Range: 3 - 6097 mics
Median: 3 mics
Average: 4 mics

So for the direct dict call, we average about 3 mics per call, while for the Signature Wrapper we average around 4. Thats a 25% cost for Signature Wrappers in this example, for a very small number of calls. Depending on what you are doing that is going to be greater or lesser. In any case, we can see that there is some cost associated with the Signature Wrapper Implementations.

What about native Signatures though? Lets take a look at ec_rbdict. The ec_rbdict also implements the ec_dictionary Signature, but it is not a Signature Wrapper. It is a native implementation of the Signature. To use ec_rbdict directly we have to create a creation helper just like we did for dict.

create_rbdict() ->
lists:foldl(fun(El, Dict) ->
        ec_rbdict:add(El, El, Dict)
    end, ec_rbdict:new(),

This is exactly the same as create_dict with the exception that dict is replaced by ec_rbdict.

The timing function itself looks very similar as well. Again notice that we have to hard code the concrete name for the concrete
implementation, but we don’t for the ec_dictionary test.

time_direct_vs_signature_rbdict() ->
    io:format("Timing rbdict~n"),
    Dict = create_rbdict(),
    test_avg(fun() ->
         ec_rbdict:size(ec_rbdict:add(some_key, some_value, Dict))
   io:format("Timing ec_dict implementation of ec_dictionary~n"),

And there we have our test. What do the results look like?

ec_dict vs ec_rbdict as an ec_dictionary Results

The main thing we are timing here is the additional cost of the dictionary Signature itself. Keep that in mind as we look at the

Erlang R14B01 (erts-5.8.2) [source] [64-bit] [smp:4:4] [rq:4] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.8.2  (abort with ^G)

1> ec_timing:time_direct_vs_signature_rbdict().
Timing rbdict
Range: 6 - 15070 mics
Median: 7 mics
Average: 7 mics
Timing ec_dict implementation of ec_dictionary
Testing ec_rbdict
Range: 6 - 6013 mics
Median: 7 mics
Average: 7 mics

So no difference it time. Well the reality is that there is a difference in timing, there must be, but we don’t have enough resolution in the timing system to be able to figure out what that difference is. Essentially that means it’s really, really small – or small enough not to worry about at the very least.


Signatures are a viable, useful approach to the problem of interfaces in Erlang. The have little or no over head depending on the type of implementation, and greatly increase the flexibility of the a library while retaining testability and locality.


: A normal Erlang Behaviour that defines a contract

: A combination of an Behaviour and functionality to make the functions callable in a concrete way

Native Signature Implementation
: A module that implements a signature directly

Signature Wrapper
: A module that does translation between a preexisting module and a Signature, allowing the preexisting module to be used as a Signature Implementation.

Code Referenced